inst/examples/swissmetro/plot_b05a_normal_mixture.R

In rbiogeme: Interface to 'Biogeme'

#!/usr/bin/env Rscript

# b05a. Normal mixture with Monte Carlo integration
#
# This example estimates a random time coefficient. Conditional on a normal
# draw, the model is a native logit kernel; native Biogeme then integrates the
# kernel with Monte Carlo draws. No R callback is used during likelihood,
# gradient, Hessian, or optimizer evaluations.

library(rbiogeme)

# The shared helper contains command-line parsing and data preparation. The
# complete model specification remains in this script.
script_path <- commandArgs(trailingOnly = FALSE)
script_path <- sub("^--file=", "", script_path[startsWith(script_path, "--file=")][[1L]])
source(file.path(dirname(normalizePath(script_path)), "example_utils.R"))

build_b05a_normal_mixture_model <- function(
    database,
    number_of_draws = 10000L,
    seed = 1223L
) {
  # These are ordinary native Biogeme parameters. The Swissmetro ASC is
  # fixed at zero to identify the utility scale.
  asc_car <- biogeme_beta("asc_car", start = 0)
  asc_train <- biogeme_beta("asc_train", start = 0)
  asc_sm <- biogeme_beta("asc_sm", start = 0, fixed = TRUE)
  b_cost <- biogeme_beta("b_cost", start = 0)
  b_time <- biogeme_beta("b_time", start = 0)
  b_time_s <- biogeme_beta("b_time_s", start = 1)

  # draw() creates a named native Draws node. The random coefficient is
  # b_time + b_time_s * NORMAL draw, with the same names as Python.
  b_time_rnd <- b_time + b_time_s * draw("b_time_rnd", "NORMAL")

  # The utilities remain symbolic R expressions. Their random coefficient is
  # compiled once into native Biogeme before numerical work begins.
  utilities <- list(
    `1` = asc_train + b_time_rnd * variable("TRAIN_TT_SCALED") +
      b_cost * variable("TRAIN_COST_SCALED"),
    `2` = asc_sm + b_time_rnd * variable("SM_TT_SCALED") +
      b_cost * variable("SM_COST_SCALED"),
    `3` = asc_car + b_time_rnd * variable("CAR_TT_SCALED") +
      b_cost * variable("CAR_CO_SCALED")
  )
  availability <- list(
    `1` = variable("TRAIN_AV_SP"),
    `2` = variable("SM_AV"),
    `3` = variable("CAR_AV_SP")
  )

  # The symbolic alternative selector is the observed CHOICE variable. This
  # compiles to native models.logit(..., i=CHOICE), rather than evaluating a
  # probability table in R. Monte Carlo integration is also native.
  kernel_probability <- logit_probability(
    utilities = utilities,
    availability = availability,
    alternative = variable("CHOICE")
  )
  log_likelihood <- log(monte_carlo(kernel_probability))
  draws <- biogeme_draws(
    name = "b_time_rnd",
    draw_type = "NORMAL",
    number_of_draws = number_of_draws,
    seed = seed
  )

  model <- biogeme_model(
    database = database,
    formula = log_likelihood,
    draws = draws
  )
  model$draws <- draws
  model
}

# prepare_swissmetro_example() is defined in example_utils.R. It parses the
# command line, validates the data/Python paths, configures the bridge, reads
# the data, and creates a fresh output directory. The --data, --python,
# --output, --draws, and --seed options work from any current working
# directory.
prepared <- prepare_swissmetro_example(
  commandArgs(trailingOnly = TRUE),
  default_model = "b05a_normal_mixture"
)

number_of_draws <- if (!is.null(prepared$options$draws) && nzchar(prepared$options$draws)) {
  example_integer(prepared$options$draws, "draws")
} else {
  10000L
}
seed <- if (!is.null(prepared$options$seed) && nzchar(prepared$options$seed)) {
  example_integer(prepared$options$seed, "seed")
} else {
  1223L
}

# estimate() always performs fresh native estimation. Remove only exact b05a
# artifacts so a reused output directory cannot silently recycle old results.
stale_files <- c(
  "b05a_normal_mixture.yaml",
  "__b05a_normal_mixture.iter",
  "b05a_normal_mixture.html"
)
stale_files <- file.path(prepared$output, stale_files)
stale_files <- stale_files[file.exists(stale_files)]
if (length(stale_files) > 0L) unlink(stale_files, force = TRUE)

database <- swissmetro_data(prepared$data)
model <- build_b05a_normal_mixture_model(
  database,
  number_of_draws = number_of_draws,
  seed = seed
)
control <- biogeme_control(
    output_directory = prepared$output,
  model_name = "b05a_normal_mixture",
  user_notes = paste0(
    "Example of a mixture of logit models with three alternatives, ",
    "approximated using Monte-Carlo integration."
  ),
  number_of_draws = number_of_draws,
  seed = seed,
  analytical_hessian_mode = "automatic",
  generate_html = TRUE,
  generate_yaml = FALSE,
  save_iterations = FALSE
)
model$control <- control

cat(sprintf("Number of draws: %s\n", format(number_of_draws, big.mark = "_")))

# The complete draw-aware expression graph is compiled once. Native Python
# Biogeme performs Monte Carlo integration and estimation.
fit <- estimate(model, model_name = "b05a_normal_mixture", control = control)

print(summary(fit))
print(coef(fit))
invisible(fit)